Commands

Supported Redis commands, Valkey compatibility, and TTL behaviour.

Current surface

Area	Commands
Connection	`PING`, `ECHO`, `SELECT 0`
Strings	`SET` (NX/XX/EX/PX), `SETNX`, `SETEX`/`PSETEX`, `GET`, `DEL`, `EXISTS`, `STRLEN`, `MGET`/`MSET`/`MSETNX`, `GETSET`, `APPEND`, `GETRANGE`/`SETRANGE`
TTL	`EXPIRE`/`PEXPIRE`, `TTL`/`PTTL`
Admin	`FLUSHDB`/`FLUSHALL`, `INFO`
Harness stubs	`CONFIG`, `FUNCTION FLUSH`

INFO returns server/memory/stats/keyspace plus pg_redis_* metadata; redis_version matches the Valkey compat pin.

Valkey compatibility

Pinned suite: Valkey 8.1.4 under compat/valkey/.

make compat   # Postgres up with in-process RESP gateway + allowlisted TCL

make compat passes --host/--port into Valkey’s harness (“external server” mode). That does not mean the standalone redis_gateway binary — it means “use the already-running embedded gateway on :6379” from make postgres-up.

Live status lives in compat/STATUS.md. Grow compat/allowlist.txt as commands land.

What the suite covers

Covered	Not covered
Allowlisted TCL tests via Valkey’s RESP client	Inline wire form (use `make bench` / raw-byte tests)
Embedded gateway + `redis.*` storage	`redis-benchmark` throughput / `PING_INLINE`
Empty-string SET/GET; TTL (lazy + `pg_cron` sweeper)	Full string suite leftovers

Green compat is necessary for semantic claims; it is not sufficient for inline protocol or benchmark defaults.

Protocol surfaces

Clients do not all speak the same wire form:

Surface	Example	Who uses it
RESP array (mbulk)	`*1\r\n$4\r\nPING\r\n`	`redis-cli`, most libraries, Valkey TCL client
Inline / telnet	`PING\r\n`	`redis-benchmark` `ping` / `PING_INLINE`, telnet
Pipelined	many commands back-to-back	`redis-benchmark -P`, busy clients

make compat and redis-cli exercise RESP arrays. They will not catch an inline-protocol regression. Default make bench uses redis-benchmark -t ping, which runs both PING_INLINE and PING_MBULK.

TTL / expiry

Lazy — expires_at columns + redis.ensure_alive on read/write.
Active (SQL modes) — pg_cron job pg_redis_ttl_task (* * * * *) calls redis.delete_expired_keys().
Active (memory mode) — 1s in-gateway tick calls MemoryStore::purge_expired().

Known limitation — naive active expiry

Path	What we do now	Gap vs Redis
Memory	Full `HashMap` scan every 1s	Redis samples a small random subset per tick
SQL (`pg_cron`)	Delete up to `batch × rounds` rows per minute	Coarser than Redis sampling cadence

SQL sweeper GUCs:

GUC	Default	Meaning
`pg_redis.ttl_delete_batch_size`	`1000`	Keys per DELETE round
`pg_redis.ttl_delete_max_rounds`	`10`	Rounds per `CALL redis.delete_expired_keys()`

Correctness (lazy-on-access + eventual active delete) is fine. Cost under large key counts is not.